Is there a standard way to compute the Normalized Point Mutual Information (NPMI) metric on topics from a topic modeling in terms of size of the sliding window of words? (in packages or in papers). I have seen mainly papers using a window of 5 words, but was there any evaluation of various window sizes? 10 words? 25 words? What about sentences or paragraphs rather than sliding windows?

More Normand Peladeau's questions See All
Similar questions and discussions